dag_ml_core/

metrics.rs

use std::collections::{BTreeMap, BTreeSet};

use serde::{Deserialize, Serialize};

use crate::aggregation::{
    reduce_predictions_across_folds, AggregatedPredictionBlock, PredictionUnitId,
};
use crate::error::{DagMlError, Result};
use crate::fold::FoldPartitionMode;
use crate::ids::{FoldId, NodeId, SampleId, VariantId};
use crate::oof::{validate_producer_oof_coverage, PredictionBlock, PredictionPartition};
use crate::policy::PredictionLevel;
use crate::selection::{CandidateScore, MetricObjective};

#[derive(Clone, Copy, Debug, Eq, PartialEq, Ord, PartialOrd, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum RegressionMetricKind {
    Mse,
    Rmse,
    Mae,
    R2,
    /// Classification accuracy: fraction of predictions whose label matches the target (integer
    /// label encoding, matched within 0.5). Meaningless on continuous regression targets (≈0) but
    /// always emitted so the host can score classification natively without a separate code path.
    Accuracy,
    /// Balanced classification accuracy: the macro-average of per-class recall (mean over the
    /// classes *present in `y_true`* of `correct_in_class / count_in_class`), matching scikit-learn's
    /// `balanced_accuracy_score`. This is nirs4all's DEFAULT classification ranking metric (its
    /// `_resolve_effective_metric` returns `balanced_accuracy` for a classification candidate), so it
    /// must be emitted natively for the dag-ml engine to reproduce the legacy classification
    /// `cv_best_score`. On a class-collapsed predictor it can be far below plain `accuracy`; on a
    /// continuous regression target it is meaningless (≈ chance) but always emitted, like `accuracy`.
    BalancedAccuracy,
}

impl RegressionMetricKind {
    pub fn name(self) -> &'static str {
        match self {
            Self::Mse => "mse",
            Self::Rmse => "rmse",
            Self::Mae => "mae",
            Self::R2 => "r2",
            Self::Accuracy => "accuracy",
            Self::BalancedAccuracy => "balanced_accuracy",
        }
    }

    pub fn objective(self) -> MetricObjective {
        match self {
            Self::Mse | Self::Rmse | Self::Mae => MetricObjective::Minimize,
            Self::R2 | Self::Accuracy | Self::BalancedAccuracy => MetricObjective::Maximize,
        }
    }
}

#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub struct RegressionTargetBlock {
    pub level: PredictionLevel,
    pub unit_ids: Vec<PredictionUnitId>,
    pub values: Vec<Vec<f64>>,
    #[serde(default)]
    pub target_names: Vec<String>,
}

impl RegressionTargetBlock {
    pub fn validate_shape(&self) -> Result<usize> {
        if self.unit_ids.len() != self.values.len() {
            return Err(DagMlError::OofValidation(format!(
                "target block has {} unit ids but {} target rows",
                self.unit_ids.len(),
                self.values.len()
            )));
        }
        if self
            .unit_ids
            .iter()
            .any(|unit_id| unit_id.level() != self.level)
        {
            return Err(DagMlError::OofValidation(format!(
                "target block contains units outside level {:?}",
                self.level
            )));
        }
        let unique = self.unit_ids.iter().collect::<BTreeSet<_>>();
        if unique.len() != self.unit_ids.len() {
            return Err(DagMlError::OofValidation(
                "target block contains duplicate unit ids".to_string(),
            ));
        }
        let width = self.values.first().map_or(0, Vec::len);
        if width == 0 {
            return Err(DagMlError::OofValidation(
                "target block has empty target rows".to_string(),
            ));
        }
        if self.values.iter().any(|row| row.len() != width) {
            return Err(DagMlError::OofValidation(
                "target block has ragged target rows".to_string(),
            ));
        }
        if self.values.iter().flatten().any(|value| !value.is_finite()) {
            return Err(DagMlError::OofValidation(
                "target block contains non-finite values".to_string(),
            ));
        }
        if !self.target_names.is_empty() && self.target_names.len() != width {
            return Err(DagMlError::OofValidation(format!(
                "target block has {} target names for width {}",
                self.target_names.len(),
                width
            )));
        }
        Ok(width)
    }
}

/// Mandatory, central *merge target-coverage* invariant — the single gate every merge reassembly
/// handler (separation/concat, fusion, off-fold) must pass through before emitting (or declining to
/// emit) a producer-level [`RegressionTargetBlock`] for a reassembled merge prediction. It closes
/// audit R-P1-9: a merge that *should* be scored must never silently produce **no** score.
///
/// A merge reassembles its output's `y_true` by collecting the per-branch validation/off-fold target
/// records into `by_sample_target`. The scoring path ([`super::runtime`]'s `apply_result_scoring`)
/// pairs a prediction block 1:1 with a target block that covers *exactly* its samples — so a partial
/// target block cannot be scored. There are exactly three legitimate outcomes:
///
/// 1. **No contributing branch emitted targets** (`by_sample_target` empty) — the merge is simply
///    unscored. Returns `Ok(None)`; the caller emits no [`RegressionTargetBlock`]. This is the common
///    "host never emitted `regression_targets`" case and stays a no-op (unchanged behavior).
/// 2. **Every merged sample is covered** — emit the 1:1 target block. Returns `Ok(Some(block))` in the
///    merge's declared `sample_id` order, ready to score.
/// 3. **At least one branch emitted targets but coverage is INCOMPLETE** — previously the partial
///    targets were silently dropped (`Vec::new()` → no score), so a merge that should have been scored
///    silently vanished from selection. This is now a hard validation **ERROR**: once ANY branch
///    contributes targets, the merge universe must be covered completely or the run fails loudly.
///
/// `merge_sample_ids` is the merge output's sample order (the universe to cover); `target_names` is the
/// reassembled target name vector (already unified across branches by the caller). The map is consumed
/// by `remove` on the success path so the caller need not clone it.
pub fn reassemble_merge_targets(
    producer_node: &NodeId,
    merge_sample_ids: &[SampleId],
    by_sample_target: &mut BTreeMap<SampleId, Vec<f64>>,
    target_names: Vec<String>,
) -> Result<Option<RegressionTargetBlock>> {
    if by_sample_target.is_empty() {
        return Ok(None);
    }
    let missing: Vec<String> = merge_sample_ids
        .iter()
        .filter(|sample_id| !by_sample_target.contains_key(*sample_id))
        .map(ToString::to_string)
        .collect();
    if !missing.is_empty() {
        return Err(DagMlError::OofValidation(format!(
            "merge node `{producer_node}` has partial target coverage: {} of {} merged sample(s) lack a y_true row ({}) while other contributing branch(es) emitted targets — a merge that some branch scores must have COMPLETE target coverage across the merge universe, never a silent no-score",
            missing.len(),
            merge_sample_ids.len(),
            missing.join(", ")
        )));
    }
    let values: Vec<Vec<f64>> = merge_sample_ids
        .iter()
        .map(|sample_id| {
            by_sample_target
                .remove(sample_id)
                .expect("target coverage was just verified complete")
        })
        .collect();
    Ok(Some(RegressionTargetBlock {
        level: PredictionLevel::Sample,
        unit_ids: merge_sample_ids
            .iter()
            .cloned()
            .map(PredictionUnitId::Sample)
            .collect(),
        values,
        target_names,
    }))
}

#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub struct RegressionMetricReport {
    #[serde(default)]
    pub prediction_id: Option<String>,
    pub producer_node: NodeId,
    /// Variant this score belongs to — distinguishes per-variant candidates when a generated
    /// campaign scores several variants, so native SELECT can pick the best. Skipped (None) for
    /// single-variant runs, so existing fixtures/fingerprints are byte-identical.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub variant_id: Option<VariantId>,
    /// Cross-language CONTENT fingerprint (hex sha256) of the operator-variant this score belongs to:
    /// the canonical form of the variant's LOWERED operator sub-sequence (Phase 5). The nirs4all host
    /// recomputes the SAME bytes from its own operator-choice config, so it can map a per-variant
    /// dag-ml report back to the config that produced it (replacing a brittle positional zip). Set
    /// only for operator-SELECT reports (the choice fingerprint from `OperatorVariantModel`'s
    /// `variant_labels`); skipped (None) for param-variant / single-variant runs, so existing
    /// fixtures/fingerprints stay byte-identical.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub variant_label: Option<String>,
    pub partition: PredictionPartition,
    pub fold_id: Option<FoldId>,
    pub level: PredictionLevel,
    pub row_count: usize,
    pub target_width: usize,
    #[serde(default)]
    pub target_names: Vec<String>,
    pub metrics: BTreeMap<String, f64>,
}

impl RegressionMetricReport {
    pub fn validate(&self) -> Result<()> {
        if self.row_count == 0 {
            return Err(DagMlError::OofValidation(
                "regression metric report has zero rows".to_string(),
            ));
        }
        if self.target_width == 0 {
            return Err(DagMlError::OofValidation(
                "regression metric report has zero target width".to_string(),
            ));
        }
        if !self.target_names.is_empty() && self.target_names.len() != self.target_width {
            return Err(DagMlError::OofValidation(format!(
                "regression metric report has {} target names for width {}",
                self.target_names.len(),
                self.target_width
            )));
        }
        if self.metrics.is_empty() {
            return Err(DagMlError::OofValidation(
                "regression metric report has no metrics".to_string(),
            ));
        }
        for (name, value) in &self.metrics {
            if name.trim().is_empty() {
                return Err(DagMlError::OofValidation(
                    "regression metric report contains an empty metric name".to_string(),
                ));
            }
            if !value.is_finite() {
                return Err(DagMlError::OofValidation(format!(
                    "regression metric `{name}` is not finite"
                )));
            }
        }
        Ok(())
    }

    pub fn into_candidate_score(self, candidate_id: impl Into<String>) -> Result<CandidateScore> {
        self.validate()?;
        let mut metadata = BTreeMap::from([
            (
                "producer_node".to_string(),
                serde_json::json!(self.producer_node),
            ),
            ("partition".to_string(), serde_json::json!(self.partition)),
            (
                "metric_level".to_string(),
                serde_json::json!(prediction_level_name(self.level)),
            ),
            ("row_count".to_string(), serde_json::json!(self.row_count)),
            (
                "target_width".to_string(),
                serde_json::json!(self.target_width),
            ),
        ]);
        if let Some(prediction_id) = self.prediction_id {
            metadata.insert(
                "prediction_id".to_string(),
                serde_json::json!(prediction_id),
            );
        }
        if let Some(fold_id) = self.fold_id {
            metadata.insert("fold_id".to_string(), serde_json::json!(fold_id));
        }
        if let Some(variant_id) = self.variant_id {
            metadata.insert("variant_id".to_string(), serde_json::json!(variant_id));
        }
        if !self.target_names.is_empty() {
            metadata.insert(
                "target_names".to_string(),
                serde_json::json!(self.target_names),
            );
        }
        let score = CandidateScore {
            candidate_id: candidate_id.into(),
            metrics: self.metrics,
            metadata,
        };
        score.validate()?;
        Ok(score)
    }
}

pub fn regression_report_to_candidate_score(
    candidate_id: impl Into<String>,
    report: RegressionMetricReport,
) -> Result<CandidateScore> {
    report.into_candidate_score(candidate_id)
}

pub fn score_regression_prediction_block(
    predictions: &PredictionBlock,
    targets: &RegressionTargetBlock,
    metrics: &[RegressionMetricKind],
) -> Result<RegressionMetricReport> {
    let width = validate_sample_prediction_block(predictions)?;
    let prediction_units = predictions
        .sample_ids
        .iter()
        .cloned()
        .map(PredictionUnitId::Sample)
        .collect::<Vec<_>>();
    score_regression_rows(
        PredictionRows {
            level: PredictionLevel::Sample,
            unit_ids: &prediction_units,
            values: &predictions.values,
            target_names: &predictions.target_names,
            width,
            origin: PredictionReportOrigin {
                prediction_id: predictions.prediction_id.clone(),
                producer_node: predictions.producer_node.clone(),
                partition: predictions.partition.clone(),
                fold_id: predictions.fold_id.clone(),
            },
        },
        targets,
        metrics,
    )
}

pub fn score_regression_aggregated_block(
    predictions: &AggregatedPredictionBlock,
    targets: &RegressionTargetBlock,
    metrics: &[RegressionMetricKind],
) -> Result<RegressionMetricReport> {
    let width = predictions.validate_shape()?;
    score_regression_rows(
        PredictionRows {
            level: predictions.level,
            unit_ids: &predictions.unit_ids,
            values: &predictions.values,
            target_names: &predictions.target_names,
            width,
            origin: PredictionReportOrigin {
                prediction_id: predictions.prediction_id.clone(),
                producer_node: predictions.producer_node.clone(),
                partition: predictions.partition.clone(),
                fold_id: predictions.fold_id.clone(),
            },
        },
        targets,
        metrics,
    )
}

/// Current on-disk schema version for [`ScoreSet`].
pub const SCORE_SET_SCHEMA_VERSION: u32 = 1;

fn default_score_set_schema_version() -> u32 {
    SCORE_SET_SCHEMA_VERSION
}

/// A persisted collection of per-block regression metric reports — the native, cross-language
/// score record produced by a run (one report per `(producer_node, partition, fold_id, level)`).
///
/// Unlike the prediction-*value* cache (which is Validation-only and leakage-gated), scores are
/// scalars derived from `y_true` and carry no feature data, so they are safe to persist for
/// every partition (train / validation / test / final). This is the score half of "dag-ml owns
/// prediction/score persistence natively" — the Python (or any host) `RunResult` reads these
/// scalars by identity, with no recomputation.
/// Identity of a score report within a [`ScoreSet`] — unique per variant, partition, fold and level.
type ScoreReportKey = (
    NodeId,
    Option<VariantId>,
    PredictionPartition,
    Option<FoldId>,
    PredictionLevel,
);

#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub struct ScoreSet {
    #[serde(default = "default_score_set_schema_version")]
    pub schema_version: u32,
    pub plan_id: String,
    /// The metric SELECT optimized for this run (e.g. `"rmse"`), if a selection ran. Metadata only.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub selection_metric: Option<String>,
    pub reports: Vec<RegressionMetricReport>,
}

impl ScoreSet {
    /// Validate the version, plan id, every report, and report-key uniqueness.
    pub fn validate(&self) -> Result<()> {
        if self.schema_version == 0 || self.schema_version > SCORE_SET_SCHEMA_VERSION {
            return Err(DagMlError::OofValidation(format!(
                "score set schema version {} is unsupported (current {SCORE_SET_SCHEMA_VERSION})",
                self.schema_version
            )));
        }
        if self.plan_id.trim().is_empty() {
            return Err(DagMlError::OofValidation(
                "score set has an empty plan_id".to_string(),
            ));
        }
        let mut seen: BTreeSet<ScoreReportKey> = BTreeSet::new();
        for report in &self.reports {
            report.validate()?;
            let key = (
                report.producer_node.clone(),
                report.variant_id.clone(),
                report.partition.clone(),
                report.fold_id.clone(),
                report.level,
            );
            if !seen.insert(key) {
                return Err(DagMlError::OofValidation(format!(
                    "score set has a duplicate report for node `{}` partition {:?} fold {:?} level {:?}",
                    report.producer_node, report.partition, report.fold_id, report.level
                )));
            }
        }
        Ok(())
    }
}

#[derive(Clone, Debug)]
struct PredictionReportOrigin {
    prediction_id: Option<String>,
    producer_node: NodeId,
    partition: PredictionPartition,
    fold_id: Option<FoldId>,
}

#[derive(Clone, Debug)]
struct PredictionRows<'a> {
    level: PredictionLevel,
    unit_ids: &'a [PredictionUnitId],
    values: &'a [Vec<f64>],
    target_names: &'a [String],
    width: usize,
    origin: PredictionReportOrigin,
}

fn score_regression_rows(
    predictions: PredictionRows<'_>,
    targets: &RegressionTargetBlock,
    metrics: &[RegressionMetricKind],
) -> Result<RegressionMetricReport> {
    if metrics.is_empty() {
        return Err(DagMlError::OofValidation(
            "no regression metrics requested".to_string(),
        ));
    }
    let mut requested_metrics = BTreeSet::new();
    for metric in metrics {
        if !requested_metrics.insert(*metric) {
            return Err(DagMlError::OofValidation(format!(
                "duplicate regression metric `{}` requested",
                metric.name()
            )));
        }
    }

    let target_width = targets.validate_shape()?;
    if predictions.width != target_width {
        return Err(DagMlError::OofValidation(format!(
            "prediction width {} does not match target width {target_width}",
            predictions.width
        )));
    }
    if predictions.level != targets.level {
        return Err(DagMlError::OofValidation(format!(
            "prediction level {:?} does not match target level {:?}",
            predictions.level, targets.level
        )));
    }
    if !predictions.target_names.is_empty()
        && !targets.target_names.is_empty()
        && predictions.target_names != targets.target_names
    {
        return Err(DagMlError::OofValidation(
            "prediction target names do not match target block names".to_string(),
        ));
    }

    let target_by_unit = targets
        .unit_ids
        .iter()
        .zip(targets.values.iter().map(Vec::as_slice))
        .collect::<BTreeMap<_, _>>();
    let mut aligned_predictions = Vec::with_capacity(predictions.unit_ids.len());
    let mut aligned_targets = Vec::with_capacity(predictions.unit_ids.len());
    for (unit_id, prediction_row) in predictions.unit_ids.iter().zip(predictions.values.iter()) {
        let target_row = target_by_unit.get(unit_id).ok_or_else(|| {
            DagMlError::OofValidation(format!(
                "prediction unit `{unit_id}` is missing from target block"
            ))
        })?;
        aligned_predictions.push(prediction_row.as_slice());
        aligned_targets.push(*target_row);
    }
    if aligned_predictions.len() != target_by_unit.len() {
        return Err(DagMlError::OofValidation(
            "target block contains units not present in predictions".to_string(),
        ));
    }

    let target_names = if !predictions.target_names.is_empty() {
        predictions.target_names.to_vec()
    } else {
        targets.target_names.clone()
    };
    let metric_suffixes = target_metric_names(predictions.width, &target_names);
    let mut values = BTreeMap::new();
    for metric in metrics {
        let per_target = compute_metric_per_target(
            *metric,
            predictions.width,
            &aligned_predictions,
            &aligned_targets,
        );
        values.insert(metric.name().to_string(), macro_mean(&per_target));
        for (name, value) in metric_suffixes.iter().zip(per_target) {
            values.insert(format!("{}:{name}", metric.name()), value);
        }
    }

    let report = RegressionMetricReport {
        prediction_id: predictions.origin.prediction_id,
        producer_node: predictions.origin.producer_node,
        variant_id: None,
        variant_label: None,
        partition: predictions.origin.partition,
        fold_id: predictions.origin.fold_id,
        level: predictions.level,
        row_count: predictions.unit_ids.len(),
        target_width: predictions.width,
        target_names,
        metrics: values,
    };
    report.validate()?;
    Ok(report)
}

fn validate_sample_prediction_block(block: &PredictionBlock) -> Result<usize> {
    block.validate_content()
}

fn compute_metric_per_target(
    metric: RegressionMetricKind,
    width: usize,
    predictions: &[&[f64]],
    targets: &[&[f64]],
) -> Vec<f64> {
    (0..width)
        .map(|target_idx| match metric {
            RegressionMetricKind::Mse => {
                predictions
                    .iter()
                    .zip(targets.iter())
                    .map(|(prediction, target)| {
                        let error = prediction[target_idx] - target[target_idx];
                        error * error
                    })
                    .sum::<f64>()
                    / predictions.len() as f64
            }
            RegressionMetricKind::Rmse => (predictions
                .iter()
                .zip(targets.iter())
                .map(|(prediction, target)| {
                    let error = prediction[target_idx] - target[target_idx];
                    error * error
                })
                .sum::<f64>()
                / predictions.len() as f64)
                .sqrt(),
            RegressionMetricKind::Mae => {
                predictions
                    .iter()
                    .zip(targets.iter())
                    .map(|(prediction, target)| (prediction[target_idx] - target[target_idx]).abs())
                    .sum::<f64>()
                    / predictions.len() as f64
            }
            RegressionMetricKind::R2 => r2_for_target(target_idx, predictions, targets),
            RegressionMetricKind::Accuracy => {
                predictions
                    .iter()
                    .zip(targets.iter())
                    .filter(|(prediction, target)| {
                        (prediction[target_idx] - target[target_idx]).abs() < 0.5
                    })
                    .count() as f64
                    / predictions.len() as f64
            }
            RegressionMetricKind::BalancedAccuracy => {
                balanced_accuracy_for_target(target_idx, predictions, targets)
            }
        })
        .collect()
}

/// Balanced classification accuracy for one target column: the macro-average of per-class recall over
/// the integer labels present in `y_true`, matching scikit-learn's `balanced_accuracy_score`. Labels
/// are matched the same way as [`RegressionMetricKind::Accuracy`] — a prediction counts for true class
/// `c` when `|pred - c| < 0.5` — so the two metrics share one label-encoding convention. Returns the
/// unweighted mean of `correct_in_class / count_in_class`; an empty target set yields `0.0` (the rows
/// are non-empty here because the scoring path rejects zero-row blocks before this is reached).
fn balanced_accuracy_for_target(
    target_idx: usize,
    predictions: &[&[f64]],
    targets: &[&[f64]],
) -> f64 {
    // Group sample rows by their (rounded) true class label, preserving determinism via BTreeMap.
    let mut per_class: BTreeMap<i64, (usize, usize)> = BTreeMap::new();
    for (prediction, target) in predictions.iter().zip(targets.iter()) {
        let true_value = target[target_idx];
        let class = true_value.round() as i64;
        let entry = per_class.entry(class).or_insert((0, 0));
        entry.1 += 1;
        if (prediction[target_idx] - true_value).abs() < 0.5 {
            entry.0 += 1;
        }
    }
    if per_class.is_empty() {
        return 0.0;
    }
    let recall_sum: f64 = per_class
        .values()
        .map(|(correct, count)| *correct as f64 / *count as f64)
        .sum();
    recall_sum / per_class.len() as f64
}

fn r2_for_target(target_idx: usize, predictions: &[&[f64]], targets: &[&[f64]]) -> f64 {
    let mean = targets.iter().map(|row| row[target_idx]).sum::<f64>() / targets.len() as f64;
    let ss_res = predictions
        .iter()
        .zip(targets.iter())
        .map(|(prediction, target)| {
            let error = prediction[target_idx] - target[target_idx];
            error * error
        })
        .sum::<f64>();
    let ss_tot = targets
        .iter()
        .map(|target| {
            let centered = target[target_idx] - mean;
            centered * centered
        })
        .sum::<f64>();
    if ss_tot == 0.0 {
        if ss_res == 0.0 {
            1.0
        } else {
            0.0
        }
    } else {
        1.0 - ss_res / ss_tot
    }
}

fn macro_mean(values: &[f64]) -> f64 {
    values.iter().sum::<f64>() / values.len() as f64
}

fn target_metric_names(width: usize, target_names: &[String]) -> Vec<String> {
    if target_names.is_empty() {
        (0..width).map(|idx| format!("target_{idx}")).collect()
    } else {
        target_names.to_vec()
    }
}

fn prediction_level_name(level: PredictionLevel) -> &'static str {
    match level {
        PredictionLevel::Observation => "observation",
        PredictionLevel::Sample => "sample",
        PredictionLevel::Target => "target",
        PredictionLevel::Group => "group",
    }
}

/// A host-emitted `y_true` block tagged with the prediction it scores (producer/partition/fold), so
/// the runtime can aggregate ground truth across folds to score cross-fold ensembles natively.
#[derive(Clone, Debug, PartialEq)]
pub struct RegressionTargetRecord {
    pub producer_node: NodeId,
    /// Variant that produced the scored block — lets the cross-fold OOF average be computed
    /// per-variant (for native SELECT) without tagging every PredictionBlock with a variant.
    pub variant_id: Option<VariantId>,
    pub partition: PredictionPartition,
    pub fold_id: Option<FoldId>,
    pub block: RegressionTargetBlock,
}

/// Combine a producer's per-fold VALIDATION `y_true` into one block (dedup by unit id — a sample's
/// ground truth is fold-independent), aligned to the producer's OOF samples.
///
/// Defense-in-depth (audit R-P0-1): records are grouped only by `producer_node`, but each carries a
/// `variant_id`. A sample's ground truth is variant-independent, so the same unit seen again must
/// carry the SAME `y_true`. If two records (e.g. from two variants sharing one context) disagree on a
/// unit's target, the ground truth has been mixed and the combined block would silently score against
/// a corrupted reference — that is refused rather than keeping whichever value happened to be first.
fn combine_validation_targets(
    producer: &NodeId,
    records: &[RegressionTargetRecord],
) -> Result<RegressionTargetBlock> {
    let mut seen: BTreeMap<PredictionUnitId, Vec<f64>> = BTreeMap::new();
    let mut unit_ids = Vec::new();
    let mut values = Vec::new();
    let mut target_names = Vec::new();
    for record in records {
        if &record.producer_node != producer || record.partition != PredictionPartition::Validation
        {
            continue;
        }
        if target_names.is_empty() {
            target_names = record.block.target_names.clone();
        }
        for (unit_id, row) in record.block.unit_ids.iter().zip(&record.block.values) {
            match seen.get(unit_id) {
                None => {
                    seen.insert(unit_id.clone(), row.clone());
                    unit_ids.push(unit_id.clone());
                    values.push(row.clone());
                }
                Some(existing) if existing != row => {
                    return Err(DagMlError::OofValidation(format!(
                        "producer `{producer}` has conflicting ground truth for unit `{unit_id:?}` across validation records — the y_true reference is mixed (e.g. several variants in one context); refusing to score against a corrupted reference"
                    )));
                }
                Some(_) => {}
            }
        }
    }
    Ok(RegressionTargetBlock {
        level: PredictionLevel::Sample,
        unit_ids,
        values,
        target_names,
    })
}

/// The per-sample cross-fold OOF average of one producer, surfaced alongside its scalar report so the
/// host can show each OOF sample's averaged prediction (nirs4all's `(validation, avg)` row y_pred),
/// not only the pooled scalar. The block is keyed by `producer_node` / `partition = Validation` /
/// `fold_id = "avg"` — identical to the scalar [`RegressionMetricReport`] this pairs with — and the
/// `y_true` covers exactly the block's samples (same id set), so the host pairs them by id. This is
/// REPORT-grade output: it carries no variant tag (the block has none; the variant is stamped on the
/// report downstream) and never feeds a training/feature path, so OOF/leakage invariants are
/// unaffected — it is purely the same averaged values the scalar was computed from, exposed per sample.
#[derive(Clone, Debug, PartialEq)]
pub struct OofAverageBlock {
    pub predictions: AggregatedPredictionBlock,
    pub y_true: RegressionTargetBlock,
}

/// The output of [`cross_fold_validation_reports`]: the scalar cross-fold OOF average reports (one per
/// producer, `fold_id = "avg"`) plus — purely additively — the per-sample OOF average block + `y_true`
/// each report was computed from. `reports` is byte-identical to the historical `Vec` return; callers
/// that only need the scalars read `reports` and ignore `oof_averages`.
#[derive(Clone, Debug, Default, PartialEq)]
pub struct CrossFoldValidation {
    pub reports: Vec<RegressionMetricReport>,
    pub oof_averages: Vec<OofAverageBlock>,
}

/// Score the cross-fold OOF average per producer: concatenate each producer's per-fold VALIDATION
/// predictions into one block and score it against the combined `y_true`. Yields one report per
/// producer with `fold_id = "avg"` — nirs4all's `cv_best_score` row — plus, additively, the per-sample
/// OOF average block + `y_true` each report was computed from (so the host can fill the `(validation,
/// avg)` row's per-sample y_pred, not only the scalar). The per-fold join is identity-keyed; producers
/// with a single fold are skipped (nothing to ensemble).
///
/// `partition_mode` mirrors the campaign [`FoldPartitionMode`]:
/// under `Partition` (KFold) the per-producer OOF must be unique (each sample scored exactly once);
/// under `Resampled` (ShuffleSplit / repeated CV) a sample may appear in several folds — those
/// predictions are averaged by [`reduce_predictions_across_folds`] — so the across-fold uniqueness
/// gate is relaxed accordingly.
pub fn cross_fold_validation_reports(
    prediction_blocks: &[PredictionBlock],
    target_records: &[RegressionTargetRecord],
    metrics: &[RegressionMetricKind],
    partition_mode: FoldPartitionMode,
) -> Result<CrossFoldValidation> {
    let mut producers: Vec<NodeId> = Vec::new();
    let mut by_producer: BTreeMap<NodeId, Vec<PredictionBlock>> = BTreeMap::new();
    for block in prediction_blocks {
        if block.partition != PredictionPartition::Validation {
            continue;
        }
        if !by_producer.contains_key(&block.producer_node) {
            producers.push(block.producer_node.clone());
        }
        by_producer
            .entry(block.producer_node.clone())
            .or_default()
            .push(block.clone());
    }
    let mut reports = Vec::new();
    let mut oof_averages = Vec::new();
    for producer in &producers {
        let blocks = &by_producer[producer];
        if blocks.len() < 2 {
            continue;
        }
        // Mandatory OOF coverage gate (spec rule 3), mode-aware. Under `Partition` the producer's
        // per-fold validation blocks must be UNIQUE — exactly one validation prediction per sample; a
        // sample appearing in two blocks would be a duplicated fold or — since `PredictionBlock` carries
        // no variant tag — two variants' OOF in a shared context (audit R-P0-1), and is refused (the
        // scoring-path analogue of the runtime merge handler's "mixes several variants" guard, so
        // cross-variant CV scores can NEVER mix here). Under `Resampled` (ShuffleSplit / repeated CV) a
        // sample is legitimately validated in several folds and its predictions are averaged by
        // `reduce_predictions_across_folds` below, so across-fold multiplicity is allowed; the per-block
        // within-fold uniqueness still holds via `validate_content`.
        let block_refs = blocks.iter().collect::<Vec<_>>();
        validate_producer_oof_coverage(producer, &block_refs, partition_mode, None)?;
        let targets = combine_validation_targets(producer, target_records)?;
        if targets.unit_ids.is_empty() {
            // No y_true was emitted for this producer (e.g. mock controllers) — nothing to score.
            continue;
        }
        let average = reduce_predictions_across_folds(blocks, None, "avg")?;
        // The scalar report is computed from `average` EXACTLY as before — byte-identical. The
        // additive per-sample surface below reuses the SAME `average` values and the SAME `targets`,
        // so it cannot perturb any score or `row_count`.
        reports.push(score_regression_prediction_block(
            &average, &targets, metrics,
        )?);
        oof_averages.push(oof_average_block(&average, &targets));
    }
    Ok(CrossFoldValidation {
        reports,
        oof_averages,
    })
}

/// Build the per-sample OOF average surface (block + `y_true`) from the SAME `average`
/// [`PredictionBlock`] and combined `targets` the scalar report was computed from. The block is the
/// sample-level lift of `average` (its sample ids become `Sample` unit ids, values unchanged), keyed
/// identically (producer / `Validation` / `avg`). The `y_true` is `targets` realigned to the block's
/// sample order so a host pairs y_pred ↔ y_true per sample without re-sorting; every `average` sample
/// has a y_true row because [`combine_validation_targets`] pools every per-fold validation record and
/// the OOF coverage gate guarantees each averaged sample was validated.
fn oof_average_block(
    average: &PredictionBlock,
    targets: &RegressionTargetBlock,
) -> OofAverageBlock {
    let unit_ids: Vec<PredictionUnitId> = average
        .sample_ids
        .iter()
        .cloned()
        .map(PredictionUnitId::Sample)
        .collect();
    let predictions = AggregatedPredictionBlock {
        prediction_id: None,
        producer_node: average.producer_node.clone(),
        partition: average.partition.clone(),
        fold_id: average.fold_id.clone(),
        level: PredictionLevel::Sample,
        unit_ids: unit_ids.clone(),
        values: average.values.clone(),
        target_names: average.target_names.clone(),
    };
    let target_by_unit: BTreeMap<&PredictionUnitId, &Vec<f64>> =
        targets.unit_ids.iter().zip(&targets.values).collect();
    let y_true = RegressionTargetBlock {
        level: PredictionLevel::Sample,
        unit_ids: unit_ids.clone(),
        values: unit_ids
            .iter()
            .map(|unit_id| target_by_unit[unit_id].clone())
            .collect(),
        target_names: targets.target_names.clone(),
    };
    OofAverageBlock {
        predictions,
        y_true,
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::ids::{FoldId, GroupId, NodeId, SampleId, TargetId};
    use crate::oof::PredictionPartition;

    fn sid(value: &str) -> SampleId {
        SampleId::new(value).unwrap()
    }

    fn sample_unit(value: &str) -> PredictionUnitId {
        PredictionUnitId::Sample(sid(value))
    }

    fn target_unit(value: &str) -> PredictionUnitId {
        PredictionUnitId::Target(TargetId::new(value).unwrap())
    }

    fn group_unit(value: &str) -> PredictionUnitId {
        PredictionUnitId::Group(GroupId::new(value).unwrap())
    }

    fn assert_close(left: f64, right: f64) {
        assert!((left - right).abs() < 1e-12, "expected {right}, got {left}");
    }

    #[test]
    fn metric_objectives_match_selection_direction() {
        assert_eq!(
            RegressionMetricKind::Rmse.objective(),
            MetricObjective::Minimize
        );
        assert_eq!(
            RegressionMetricKind::Mae.objective(),
            MetricObjective::Minimize
        );
        assert_eq!(
            RegressionMetricKind::Mse.objective(),
            MetricObjective::Minimize
        );
        assert_eq!(
            RegressionMetricKind::R2.objective(),
            MetricObjective::Maximize
        );
    }

    #[test]
    fn reassemble_merge_targets_empty_map_is_unscored_none() {
        // No contributing branch emitted targets: the merge is legitimately unscored.
        let producer = NodeId::new("merge:m").unwrap();
        let mut by_sample: BTreeMap<SampleId, Vec<f64>> = BTreeMap::new();
        let block = reassemble_merge_targets(
            &producer,
            &[sid("s1"), sid("s2")],
            &mut by_sample,
            vec!["y".to_string()],
        )
        .unwrap();
        assert!(
            block.is_none(),
            "empty targets -> unscored None, not an error"
        );
    }

    #[test]
    fn reassemble_merge_targets_complete_coverage_emits_ordered_block() {
        let producer = NodeId::new("merge:m").unwrap();
        let mut by_sample: BTreeMap<SampleId, Vec<f64>> = BTreeMap::new();
        by_sample.insert(sid("s2"), vec![20.0]);
        by_sample.insert(sid("s1"), vec![10.0]);
        let block = reassemble_merge_targets(
            &producer,
            &[sid("s1"), sid("s2")],
            &mut by_sample,
            vec!["y".to_string()],
        )
        .unwrap()
        .expect("complete coverage -> a target block");
        // Emitted in the merge's declared sample order, not map order.
        assert_eq!(
            block.unit_ids,
            vec![sample_unit("s1"), sample_unit("s2")],
            "targets follow the merge sample order"
        );
        assert_eq!(block.values, vec![vec![10.0], vec![20.0]]);
        assert_eq!(block.level, PredictionLevel::Sample);
        block.validate_shape().unwrap();
    }

    #[test]
    fn reassemble_merge_targets_partial_coverage_is_validation_error() {
        // R-P1-9: one branch contributed a target (s1) but the merge universe also
        // covers s2, which has no y_true row. Previously this silently dropped the
        // targets (no score); it must now be a hard validation error.
        let producer = NodeId::new("merge:m").unwrap();
        let mut by_sample: BTreeMap<SampleId, Vec<f64>> = BTreeMap::new();
        by_sample.insert(sid("s1"), vec![10.0]);
        let err = reassemble_merge_targets(
            &producer,
            &[sid("s1"), sid("s2")],
            &mut by_sample,
            vec!["y".to_string()],
        )
        .unwrap_err();
        let msg = err.to_string();
        assert!(
            msg.contains("partial target coverage") && msg.contains("s2"),
            "partial coverage names the missing sample: {msg}"
        );
    }

    #[test]
    fn scores_sample_predictions_and_exports_candidate_metrics() {
        let predictions = PredictionBlock {
            prediction_id: Some("pred:sample".to_string()),
            producer_node: NodeId::new("model:pls").unwrap(),
            partition: PredictionPartition::Validation,
            fold_id: None,
            sample_ids: vec![sid("sample:1"), sid("sample:2")],
            values: vec![vec![2.0], vec![4.0]],
            target_names: vec!["y".to_string()],
        };
        let targets = RegressionTargetBlock {
            level: PredictionLevel::Sample,
            unit_ids: vec![sample_unit("sample:2"), sample_unit("sample:1")],
            values: vec![vec![5.0], vec![1.0]],
            target_names: vec!["y".to_string()],
        };

        let report = score_regression_prediction_block(
            &predictions,
            &targets,
            &[
                RegressionMetricKind::Rmse,
                RegressionMetricKind::Mae,
                RegressionMetricKind::R2,
            ],
        )
        .unwrap();

        assert_eq!(report.level, PredictionLevel::Sample);
        assert_close(report.metrics["rmse"], 1.0);
        assert_close(report.metrics["rmse:y"], 1.0);
        assert_close(report.metrics["mae"], 1.0);
        assert_close(report.metrics["r2"], 0.75);
        let candidate = regression_report_to_candidate_score("model:pls", report).unwrap();
        assert_eq!(candidate.metrics["rmse"], 1.0);
        assert_eq!(candidate.metadata["metric_level"], "sample");
        assert_eq!(candidate.metadata["producer_node"], "model:pls");
        assert_eq!(candidate.metadata["partition"], "validation");
        assert_eq!(candidate.metadata["prediction_id"], "pred:sample");
        assert_eq!(candidate.metadata["target_names"], serde_json::json!(["y"]));
    }

    #[test]
    fn scores_target_and_group_prediction_blocks() {
        let predictions = AggregatedPredictionBlock {
            prediction_id: Some("pred:target".to_string()),
            producer_node: NodeId::new("model:pls").unwrap(),
            partition: PredictionPartition::Validation,
            fold_id: None,
            level: PredictionLevel::Target,
            unit_ids: vec![target_unit("target:a"), target_unit("target:b")],
            values: vec![vec![1.0, 10.0], vec![3.0, 30.0]],
            target_names: vec!["y1".to_string(), "y2".to_string()],
        };
        let targets = RegressionTargetBlock {
            level: PredictionLevel::Target,
            unit_ids: vec![target_unit("target:b"), target_unit("target:a")],
            values: vec![vec![2.0, 28.0], vec![2.0, 12.0]],
            target_names: vec!["y1".to_string(), "y2".to_string()],
        };
        let report = score_regression_aggregated_block(
            &predictions,
            &targets,
            &[RegressionMetricKind::Mse, RegressionMetricKind::Rmse],
        )
        .unwrap();

        assert_eq!(report.level, PredictionLevel::Target);
        assert_close(report.metrics["mse:y1"], 1.0);
        assert_close(report.metrics["mse:y2"], 4.0);
        assert_close(report.metrics["mse"], 2.5);
        assert_close(report.metrics["rmse:y1"], 1.0);
        assert_close(report.metrics["rmse:y2"], 2.0);
        assert_close(report.metrics["rmse"], 1.5);

        let group_predictions = AggregatedPredictionBlock {
            prediction_id: Some("pred:group".to_string()),
            producer_node: NodeId::new("model:pls").unwrap(),
            partition: PredictionPartition::Validation,
            fold_id: None,
            level: PredictionLevel::Group,
            unit_ids: vec![group_unit("group:a")],
            values: vec![vec![3.0]],
            target_names: vec!["y".to_string()],
        };
        let group_targets = RegressionTargetBlock {
            level: PredictionLevel::Group,
            unit_ids: vec![group_unit("group:a")],
            values: vec![vec![1.0]],
            target_names: vec!["y".to_string()],
        };
        let group_report = score_regression_aggregated_block(
            &group_predictions,
            &group_targets,
            &[RegressionMetricKind::Mae],
        )
        .unwrap();
        assert_eq!(group_report.level, PredictionLevel::Group);
        assert_close(group_report.metrics["mae"], 2.0);
    }

    #[test]
    fn refuses_metric_alignment_and_contract_mismatches() {
        let predictions = AggregatedPredictionBlock {
            prediction_id: None,
            producer_node: NodeId::new("model:pls").unwrap(),
            partition: PredictionPartition::Validation,
            fold_id: None,
            level: PredictionLevel::Target,
            unit_ids: vec![target_unit("target:a")],
            values: vec![vec![1.0]],
            target_names: vec!["y".to_string()],
        };
        let missing_target = RegressionTargetBlock {
            level: PredictionLevel::Target,
            unit_ids: vec![target_unit("target:b")],
            values: vec![vec![1.0]],
            target_names: vec!["y".to_string()],
        };
        assert!(score_regression_aggregated_block(
            &predictions,
            &missing_target,
            &[RegressionMetricKind::Rmse],
        )
        .is_err());

        let wrong_level = RegressionTargetBlock {
            level: PredictionLevel::Group,
            unit_ids: vec![group_unit("group:a")],
            values: vec![vec![1.0]],
            target_names: vec!["y".to_string()],
        };
        assert!(score_regression_aggregated_block(
            &predictions,
            &wrong_level,
            &[RegressionMetricKind::Rmse],
        )
        .is_err());

        assert!(score_regression_aggregated_block(&predictions, &missing_target, &[]).is_err());
        assert!(score_regression_aggregated_block(
            &predictions,
            &RegressionTargetBlock {
                level: PredictionLevel::Target,
                unit_ids: vec![target_unit("target:a")],
                values: vec![vec![1.0]],
                target_names: vec!["other".to_string()],
            },
            &[RegressionMetricKind::Rmse],
        )
        .is_err());
        assert!(score_regression_aggregated_block(
            &predictions,
            &RegressionTargetBlock {
                level: PredictionLevel::Target,
                unit_ids: vec![target_unit("target:a")],
                values: vec![vec![1.0]],
                target_names: vec!["y".to_string()],
            },
            &[RegressionMetricKind::Rmse, RegressionMetricKind::Rmse],
        )
        .is_err());
    }

    #[test]
    fn refuses_duplicate_and_non_finite_sample_predictions() {
        let targets = RegressionTargetBlock {
            level: PredictionLevel::Sample,
            unit_ids: vec![sample_unit("sample:1")],
            values: vec![vec![1.0]],
            target_names: vec!["y".to_string()],
        };
        let mut predictions = PredictionBlock {
            prediction_id: None,
            producer_node: NodeId::new("model:pls").unwrap(),
            partition: PredictionPartition::Validation,
            fold_id: None,
            sample_ids: vec![sid("sample:1")],
            values: vec![vec![f64::INFINITY]],
            target_names: vec!["y".to_string()],
        };
        assert!(score_regression_prediction_block(
            &predictions,
            &targets,
            &[RegressionMetricKind::Rmse],
        )
        .is_err());

        predictions.values = vec![vec![1.0], vec![1.0]];
        predictions.sample_ids = vec![sid("sample:1"), sid("sample:1")];
        assert!(score_regression_prediction_block(
            &predictions,
            &targets,
            &[RegressionMetricKind::Rmse],
        )
        .is_err());
    }

    #[test]
    fn constant_target_r2_is_finite_and_deterministic() {
        let targets = RegressionTargetBlock {
            level: PredictionLevel::Sample,
            unit_ids: vec![sample_unit("sample:1"), sample_unit("sample:2")],
            values: vec![vec![2.0], vec![2.0]],
            target_names: vec!["y".to_string()],
        };
        let exact_predictions = PredictionBlock {
            prediction_id: None,
            producer_node: NodeId::new("model:exact").unwrap(),
            partition: PredictionPartition::Validation,
            fold_id: None,
            sample_ids: vec![sid("sample:1"), sid("sample:2")],
            values: vec![vec![2.0], vec![2.0]],
            target_names: vec!["y".to_string()],
        };
        let exact_report = score_regression_prediction_block(
            &exact_predictions,
            &targets,
            &[RegressionMetricKind::R2],
        )
        .unwrap();
        assert_close(exact_report.metrics["r2"], 1.0);

        let off_predictions = PredictionBlock {
            values: vec![vec![2.0], vec![3.0]],
            ..exact_predictions
        };
        let off_report = score_regression_prediction_block(
            &off_predictions,
            &targets,
            &[RegressionMetricKind::R2],
        )
        .unwrap();
        assert_close(off_report.metrics["r2"], 0.0);
    }

    fn score_report(
        partition: PredictionPartition,
        fold: Option<&str>,
        rmse: f64,
    ) -> RegressionMetricReport {
        RegressionMetricReport {
            prediction_id: None,
            producer_node: NodeId::new("model:compat.0").unwrap(),
            variant_id: None,
            variant_label: None,
            partition,
            fold_id: fold.map(|value| FoldId::new(value).unwrap()),
            level: PredictionLevel::Sample,
            row_count: 10,
            target_width: 1,
            target_names: vec!["y".to_string()],
            metrics: BTreeMap::from([("rmse".to_string(), rmse), ("r2".to_string(), 0.5)]),
        }
    }

    #[test]
    fn score_set_round_trips_validates_and_rejects_duplicates() {
        let set = ScoreSet {
            schema_version: SCORE_SET_SCHEMA_VERSION,
            plan_id: "plan:demo".to_string(),
            selection_metric: Some("rmse".to_string()),
            reports: vec![
                score_report(PredictionPartition::Validation, Some("avg"), 18.75),
                score_report(PredictionPartition::Test, Some("final"), 13.28),
            ],
        };
        set.validate().unwrap();

        // JSON round-trip is lossless.
        let json = serde_json::to_string(&set).unwrap();
        let back: ScoreSet = serde_json::from_str(&json).unwrap();
        assert_eq!(back, set);

        // schema_version defaults when omitted (forward-compatible read).
        let parsed: ScoreSet =
            serde_json::from_value(serde_json::json!({"plan_id": "p", "reports": []})).unwrap();
        assert_eq!(parsed.schema_version, SCORE_SET_SCHEMA_VERSION);

        // Duplicate (producer_node, partition, fold_id, level) is rejected.
        let dup = ScoreSet {
            reports: vec![
                score_report(PredictionPartition::Test, Some("final"), 1.0),
                score_report(PredictionPartition::Test, Some("final"), 2.0),
            ],
            ..set.clone()
        };
        assert!(dup.validate().is_err());

        // Empty plan_id is rejected.
        let blank = ScoreSet {
            plan_id: "  ".to_string(),
            reports: vec![score_report(PredictionPartition::Test, Some("final"), 1.0)],
            ..set
        };
        assert!(blank.validate().is_err());
    }

    #[test]
    fn accuracy_and_balanced_accuracy_match_sklearn_on_imbalanced_classification() {
        // #60 root-cause lock: dag-ml emits BOTH plain `accuracy` and `balanced_accuracy`. nirs4all's
        // DEFAULT classification ranking metric is balanced_accuracy (its `_resolve_effective_metric`),
        // so the legacy `cv_best_score` for a classification sweep is balanced_accuracy — NOT plain
        // accuracy. A class-collapsed predictor on imbalanced data makes the two diverge sharply (the
        // 0.32-vs-0.16 STOP report). Ground truth here is scikit-learn on the same labels:
        //   y    = [0,0,0,0,0,0, 1,1, 2,2]  (majority class 0)
        //   pred = [0,0,0,0,0,0, 1,0, 0,0]  (collapses wrong rows to class 0)
        //   accuracy_score          = 7/10 = 0.70
        //   balanced_accuracy_score = mean(recall(c0)=6/6, recall(c1)=1/2, recall(c2)=0/2) = 0.50
        let predictions = PredictionBlock {
            prediction_id: Some("pred:classif".to_string()),
            producer_node: NodeId::new("model:rf").unwrap(),
            partition: PredictionPartition::Validation,
            fold_id: None,
            sample_ids: (0..10).map(|i| sid(&format!("s{i}"))).collect(),
            values: vec![
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![1.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
            ],
            target_names: vec!["y".to_string()],
        };
        let targets = RegressionTargetBlock {
            level: PredictionLevel::Sample,
            unit_ids: (0..10).map(|i| sample_unit(&format!("s{i}"))).collect(),
            values: vec![
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![1.0],
                vec![1.0],
                vec![2.0],
                vec![2.0],
            ],
            target_names: vec!["y".to_string()],
        };

        let report = score_regression_prediction_block(
            &predictions,
            &targets,
            &[
                RegressionMetricKind::Accuracy,
                RegressionMetricKind::BalancedAccuracy,
            ],
        )
        .unwrap();

        assert_close(report.metrics["accuracy"], 0.70);
        assert_close(report.metrics["balanced_accuracy"], 0.50);
        // Both maximize — the host SELECT ranks them in the same direction.
        assert_eq!(
            RegressionMetricKind::BalancedAccuracy.objective(),
            MetricObjective::Maximize
        );
    }

    #[test]
    fn cross_fold_balanced_accuracy_pools_oof_and_matches_sklearn() {
        // The real #60 path: per-fold VALIDATION OOF blocks pooled into the `avg` report (nirs4all's
        // `cv_best_score` row), scored against the combined y_true. Two disjoint KFold folds carry the
        // same imbalanced class-collapse as the per-block lock above, so the POOLED OOF accuracy is
        // 0.70 and pooled balanced_accuracy is 0.50 — proving the cross-fold reduction + metric agree
        // with scikit-learn's `accuracy_score` / `balanced_accuracy_score` on the same labels, and that
        // dag-ml's `accuracy` (0.70) was never wrong: it is simply a DIFFERENT metric from the legacy's
        // default `balanced_accuracy` (0.50).
        let model = NodeId::new("model:rf").unwrap();
        let fold_block = |fold: &str, ids: &[usize], preds: &[f64]| PredictionBlock {
            prediction_id: Some(format!("pred:{fold}")),
            producer_node: model.clone(),
            partition: PredictionPartition::Validation,
            fold_id: Some(FoldId::new(fold).unwrap()),
            sample_ids: ids.iter().map(|i| sid(&format!("s{i}"))).collect(),
            values: preds.iter().map(|p| vec![*p]).collect(),
            target_names: vec!["y".to_string()],
        };
        let target_record = |fold: &str, ids: &[usize], trues: &[f64]| RegressionTargetRecord {
            producer_node: model.clone(),
            variant_id: None,
            partition: PredictionPartition::Validation,
            fold_id: Some(FoldId::new(fold).unwrap()),
            block: RegressionTargetBlock {
                level: PredictionLevel::Sample,
                unit_ids: ids.iter().map(|i| sample_unit(&format!("s{i}"))).collect(),
                values: trues.iter().map(|t| vec![*t]).collect(),
                target_names: vec!["y".to_string()],
            },
        };

        // Fold 0 (samples 0..5): preds [0,0,0,1,0] vs true [0,0,0,1,2]
        // Fold 1 (samples 5..10): preds [0,0,0,0,0] vs true [0,0,0,1,2]
        // Pooled: true=[0,0,0,1,2,0,0,0,1,2], pred=[0,0,0,1,0,0,0,0,0,0]
        //   accuracy          = 7/10 = 0.70
        //   balanced_accuracy = mean(recall0=6/6, recall1=1/2, recall2=0/2) = 0.50
        let f0 = (0..5).collect::<Vec<_>>();
        let f1 = (5..10).collect::<Vec<_>>();
        let blocks = vec![
            fold_block("0", &f0, &[0.0, 0.0, 0.0, 1.0, 0.0]),
            fold_block("1", &f1, &[0.0, 0.0, 0.0, 0.0, 0.0]),
        ];
        let targets = vec![
            target_record("0", &f0, &[0.0, 0.0, 0.0, 1.0, 2.0]),
            target_record("1", &f1, &[0.0, 0.0, 0.0, 1.0, 2.0]),
        ];

        let outcome = cross_fold_validation_reports(
            &blocks,
            &targets,
            &[
                RegressionMetricKind::Accuracy,
                RegressionMetricKind::BalancedAccuracy,
            ],
            FoldPartitionMode::Partition,
        )
        .unwrap();

        assert_eq!(
            outcome.reports.len(),
            1,
            "one pooled `avg` report for the producer"
        );
        let avg = &outcome.reports[0];
        assert_eq!(avg.fold_id, Some(FoldId::new("avg").unwrap()));
        assert_eq!(avg.row_count, 10, "all OOF samples pooled exactly once");
        assert_close(avg.metrics["accuracy"], 0.70);
        assert_close(avg.metrics["balanced_accuracy"], 0.50);

        // Additive per-sample OOF average surface: one block per scored producer, keyed identically to
        // the scalar report (producer / Validation / avg), with the SAME pooled values and one y_true
        // row per averaged sample (same id set), realigned to the block's sample order.
        assert_eq!(outcome.oof_averages.len(), 1, "one OOF average block");
        let oof = &outcome.oof_averages[0];
        assert_eq!(oof.predictions.partition, PredictionPartition::Validation);
        assert_eq!(oof.predictions.fold_id, Some(FoldId::new("avg").unwrap()));
        assert_eq!(oof.predictions.level, PredictionLevel::Sample);
        assert_eq!(oof.predictions.unit_ids.len(), 10);
        assert_eq!(oof.y_true.unit_ids, oof.predictions.unit_ids);
        // The pooled per-sample preds are the across-fold mean (each KFold sample validated once):
        // [0,0,0,1,0] ++ [0,0,0,0,0] against y_true [0,0,0,1,2] ++ [0,0,0,1,2].
        assert_eq!(
            oof.predictions.values,
            vec![
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![1.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
            ]
        );
        assert_eq!(
            oof.y_true.values,
            vec![
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![1.0],
                vec![2.0],
                vec![0.0],
                vec![0.0],
                vec![0.0],
                vec![1.0],
                vec![2.0],
            ]
        );
    }
}